Discriminative Corpus Weight Estimation for Machine Translation

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 708–717, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Discriminative Corpus Weight Estimation for Machine Translation Spyros Matsoukas and Antti-Veikko I. Rosti and Bing Zhang BBN Technologies, 10 Moulton Street, Cambridge, MA 02138 {smatsouk,arosti,bzhang}@bbn.com Abstract Current statistical machine translation (SMT) systems are trained on sentence- aligned and word-aligned parallel text col- lected from various sources. Translation model parameters are estimated from the word alignments, and the quality of the translations on a given test set depends on the parameter estimates. There are at least two factors affecting the parame- ter estimation: domain match and training data quality. This paper describes a novel approach for automatically detecting and down-weighing certain parts of the train- ing corpus by assigning a weight to each sentence in the training bitext so as to op- timize a discriminative objective function on a designated tuning set. This way, the proposed method can limit the negative ef- fects of low quality training data, and can adapt the translation model to the domain of interest. It is shown that such discrim- inative corpus weights can provide sig- nificant improvements in Arabic-English translation on various conditions, using a state-of-the-art SMT system. 1 Introduction Statistical machine translation (SMT) systems rely on a training corpus consisting of sentences in the source language and their respective reference translations to the target language. These paral- lel sentences are used to perform automatic word alignment, and extract translation rules with asso- ciated probabilities. Typically, a parallel training corpus is comprised of collections of varying qual- ity and relevance to the translation problem of in- terest. For example, an SMT system applied to broadcast conversational data may be trained on a corpus consisting mostly of United Nations and newswire data, with only a very small amount of in-domain broadcast news/conversational data. In this case, it would be desirable to down-weigh the out-of-domain data relative to the in-domain data during the rule extraction and probability estima- tion. Similarly, it would be good to assign a lower weight to data of low quality (e.g., poorly aligned or incorrectly translated sentences) relative to data of high quality. In this paper, we describe a novel discrimina- tive training method that can be used to estimate a weight for each sentence in the training bitext so as to optimize an objective function – expected trans- lation edit rate (TER) (Snover et al., 2006) – on a held-out development set. The training bitext typ- ically consists of millions of (parallel) sentences, so in order to ensure robust estimation we express each sentence weight as a function of sentence- level features, and estimate the parameters of this mapping function instead. Sentence-level fea- tures may include the identifier of the collection or genre that the sentence belongs to, the number of tokens in the source or target side, alignment infor- mation, etc. The mapping from features to weights can be implemented via any differentiable func- tion, but in our experiments …

Content truncated to 500 words